VAETKI-VL 7B-A1B (GGUF)
dororodoroddo/VAETKI-VL-7B-A1B-GGUF is a GGUF packaging of nc-ai-consortium/VAETKI-VL-7B-A1B, a Korean/English vision-language model built as a Mixture-of-Experts. The model card calls out 7.58B total parameters with ~1.2B activated (64 experts, 5 active per token), plus a 16k context window — a design that aims to keep inference relatively efficient while still offering a larger-capacity backbone.
What makes this repo practically useful is that it’s structured the way the llama.cpp ecosystem expects: a quantized text model GGUF plus a separate mmproj GGUF for the vision encoder. That means it’s geared toward running locally (for example, with llama.cpp’s multimodal CLI) for tasks like image captioning, describing screenshots, or Korean/English visual Q&A.
The catch is compatibility: the author notes that llama.cpp support is pending an upstream merge and points to a fork in the meantime. If you want to try it, treat it like an “early access” integration — validate your exact llama.cpp build supports VAETKI-VL before you commit to it for anything beyond experimentation.
Quick stats from the listing feed: pipeline: image-text-to-text · 2 likes · 56 downloads.
Source listing: https://huggingface.co/models?sort=modified